Acorn Heroes

In my previous articles, we’ve looked at how we can set up an automatic build, with a focus on catching issues with bad check-ins immediately.  Another useful aspect of an automated build is that it can remove the chance of human error whenever we have a task that consists of a lot of manual steps.

One such example is generating a build for ad-hoc distribution to beta testers.  We could create a build by hand, but if we forget to increment the version number in our project, then our testers won’t be able to update from a previous build unless they delete the old build by hand off their devices and iTunes – definitely a pain for people you don’t want to inconvenience.

Also, imagine the scenario in a couple of weeks when a tester notifies you of a serious bug they’ve found.  Chances are the code you have now is quite different to the build they have.  Can you get back to that point in the code easily?

Fortunately Apple provide us with a handy tool called agvtool that can handle these situations for us.  It’s a command line tool, so it integrates nicely into Hudson.  But first let’s look at agvtool on its own, and see what we can do with it.

agvtool

In order for agvtool to do its magic, we have a few things to set up in our project first.  Most of the steps here are described in detail by Jamie Montgomerie on his blog.  Briefly though, we will use two different version numbers – a release version and a build number.  The release version is our “marketing” number – what our clients see (“Now updated to version 1.2!”).  Our build number is for internal use, and will connect a given instance of our App to a specific tagged version of our code in source control. To test this, I created a simple OpenGL project in Xcode and added it to SVN as a starting point.  The only modification I’ve made to it is to add a label on top of the GL view.  We will use this later to display our build number for testers.  Note that SVN is not necessary for using agvtool, but it will give us a few nice options down the track.

So, following Jamie’s instructions, we:

1. Open our project’s PROJECT_NAME-Info.plist file and set the Bundle version to 1.  This represents our build number.
2. Ctrl (or right) click on the plist and choose Add Row.  From the options presented, select Bundle versions string, short and set its initial value to (say) 0.1.  This is our release version.
3. Ctrl (or right) click on the project in the Groups & Files window, and select Get Info.  Make sure you’ve selected All Configurations and find the Versioning section, near the bottom.  Set the Current Project Version to match you build number – 1 in our case.  Set the Versioning system to apple-generic.

OK, not too painful and we’ll never need to touch these bits again!  Let’s test our a few things.  Open up a terminal window and cd to your project directory.  We can use agvtool to tell us both the current build and release version numbers (agvtool calls them version and marketing version respectively):

You should see the build number (agvtool what-version) and release version (agvtool what-marketing-version) that you entered into the project settings.  Now, using agvtool we can also update our build and release version numbers.  Note that doing this modifies your project, so ensure it’s either not open in XCode or everything has been saved – otherwise you may lose changes to your project such as settings or newly added files. So, to increment our build number we use:

agvtool bump -all

On the slightly rarer occasion that we want to adjust our release version, this can be done as follows:

agvtool new-marketing-version 0.21b

You can use the agvtool what- command to verify what’s going on here.  Note that our build number is automatically incremented by one using the bump command.  Our release version is in fact a string, and we can supply whatever we want there, in this case 0.21b.

Assuming you’re paranoid like me, now is a good time to check this into source control as we have all the basics covered.

Open the project in XCode and build it.  XCode will now automatically generate a .c file as part of the build.  In my case, these were found at:

build/AutoVersionTest.build/Debug-iphonesimulator/AutoVersionTest.build/DerivedSources/AutoVersionTest_vers.c

You don’t need to even open this file, or include them in your project (that happens automatically).  This file declares two variables, a version string and number that correspond to our build number.  We can put them to good use.  In my sample project, I’ve declared the build number as an external variable in my App delegate:

extern const double AutoVersionTestVersionNumber;

And then in my application: didFinishLaunchingWithOptions method I’ve used the build number on the label I added earlier:

versionLabel.text = [NSString stringWithFormat:@"Build: %1.0f",AutoVersionTestVersionNumber, nil];

Now when testers run my App they’ll see something like this:

Now, a tester can easily identify the build they’re running when giving me feedback.  By the way – if anyone know an easy way to get at our release version number in code, please let me know!  What we need now is a simple way to tie that back to a specific version of our code…

Tagging in SVN

Note, this also works with CVS as well, however while there are good arguments for not moving to a more modern system like Git or Mercurial, there’s no good argument for sticking with CVS.

agvtool has the ability to update our version information and then commit the changes, simply by adding the usesvn tag like so:

agvtool -usesvn bump -all

Nice.  But it gets better.  We can then tag these changes to make getting back to them simple:

agvtool -usesvn tag -baseurlfortag http://URL/to/svn/AutoVersionTest/tags

The -baseurlfortag option specifies where the tag should go.  After this operation, you’re SVN repository should look something like this:

Now, when you receive feedback from a tester it’s a simple matter to check out the tagged code and you can track down issues with the same code base your tester was using.  Note that the folder specified by the -baseurlfortag option must already exist – agvtool will not create it for you.

And Hudson?

Hopefully it’s not too much of a stretch from here to see how we can use this in Hudson (or any other automated build system).  First off, we should create a project in Hudson that will check out our code and build it.  If you already have an existing Hudson project that builds periodically or with each new check-in you can simply clone it (Hudson is good that way).  Configure the project to only build manually by unchecking all the build triggers:

And add in new build step(s), calling agvtool to bump our build number and (if using SVN) tag the resulting build.  Note that if this project were set up to trigger a new build when check-ins are detected we have a potential race condition, with agvtool checking in changes which would trigger a new build which would make agvtool check in changes….

So, with these changes in place we can now reliably build a version of our code for testers, tag it and know that we can come back to the code for that build at any time.  And it’s all done automatically.  Also our version numbers will always increase, insuring that our testers don’t have trouble updating to a newer build.

And Another Thing

OK, I haven’t played with this one extensively yet, but will add it here for your consideration.  Hudson allows us to add parameters to a build.  When we start a build, it will ask us to supply parameters that affect the build.  As an example we can add a BuildNumber parameter that will decide which tagged version of our code we want.  Configuring it in Hudson will look something like this:

So what’s the use of this?  Well the parameter becomes an environment variable when Hudson is building the project.  So we can slip the environment variable into our source control link.  This way, when we kick off a build we can specify the tag to use, making it easy to recreate our App as it was at any stage in its development.

Well that’s it for this week.  If you have any questions or feedback please leave a comment below.

This post is part of iDevBlogADay, a group of indie iPhone development blogs featuring two posts per day. You can keep up with iDevBlogADay through the web site, RSS feed, or Twitter.

After last week’s introduction, it’s time to get on with making this thing.  After looking at both CruiseControl and Hudson, I went with Hudson – mainly because it’s very simple to set up and start using straight away.  So without further fuss, let’s download Hudson from here.  I created a build folder ~/AutomatedBuilds (there’s nothing special about this location, put it wherever suits you) and put the downloaded file hudson.war.zip in it.

Open a terminal window and cd to the build folder.  The online documentation says that you should unzip hudson and run it using the command:

java -jar hudson.war

This works fine on Windows, however under OS X, unzipping the file and trying to run it causes the following error:

If anyone can explain the issue, I’d appreciate the insight.  However it turns out we can just run Hudson as-is out of the zip file:

Ok, so what’s happened?  Hudson is now running as a Java application in the terminal.  As well as running the build, it also provides a web interface for us to work with.  Open a browser and point it at http://localhost:8080/ and you should see something like:

Before we set up our first job, we should grab a couple of handy plugins first.  Click on the Manage Hudson link on the left, then Manage Plugins and finally the Available tab.  You should see a long, slightly exciting (or intimidating, depending on your outlook on life) list of plugins.  Have a look through the list – there’s lots here to play with – but for now select the Twitter and Mercurial plugins (assuming I’ve convinced you to try Mercurial), and click on the Install button.

You should see the plugins being downloaded and installed, along with a message to restart Hudson.  When Hudson is running, there’s a Prepare for Shutdown option on the Manage Hudson page.  For now though, it’s enough to go to the Terminal window you’re running Hudson in and kill it with Ctrl+C (I’m a bit old fashioned that way).

Run Hudson again from the terminal window, and we’re ready to set up our first automated build.  For a start, let’s create a build that monitors our source control repository and kicks off a build when it sees any changes.

On the main Hudson dashboard, click New Job.  Give the job a name, choose the “free-style software project” option and click OK.

On the resulting page you can configure a bunch of options.  Feel free to go nuts here.  Each option has a ‘?’ icon that provides useful help.  Here’s the settings I used:

Under Source Code Management, I chose Mercurial and pointed it at my Fogbugz Repository.  The Fogbugz repository requires the user to be authenticated, so we do this by including the user name and password in the URL in the form:

https://USER_NAME:PASSWORD@PATH_TO_REPOSITORY

Note that, if your user name has a ‘@’ symbol in it, replace the ‘@’ with ‘%40′ to avoid the URL being incorrectly parsed by Hudson or Fogbugz (see the screen shot below).

Hudson also supports many other types of source control, pick the one that you use.  Another small note here – when setting up an automated build for the first time, it’s a good idea to use a small test project first, to allow faster testing of your set up.  My actual production code can take many minutes to build (not my fault! OpenFeint is slow to build), so instead, for this article, I’ve created AwesomeGame from one of the standard XCode templates.  Once I have a running build system, I’ll then go back and add in my production code.

Next up, we want to check if we need to trigger a build by polling SCM for changes.  We schedule Hudson to check SCM every minute using CRON-like syntax :

Then, we add the actual build step, executing a shell command to run XCode from the command line using xcodebuild.  Have a look at the xcodebuild man page for all the various options.  For now, we want to build our AwesomeGame target in both Debug and Release modes.  I’ve done this below using two build steps:

Finally, let’s tell Hudson how to notify us when the build is done.  I selected email and Twitter notifications.  If you’re on a team, notice the option to notify the individual who committed the changes that broke the build – without having to notify the entire team.  As well as the person responsible for the break it’s good practice to have a nominated ‘build monkey’ whose job it is to check all builds – so make sure they’re getting all the build results as well.   A quick note on Twitter here – don’t use your own Twitter account unless you want to spam all your followers!  Instead, create a bogus build account for Hudson to use, and follow it from your main account.

Assuming we’re all happy with the settings (we can always come back and tweak them), click Save.

Hudson now takes us to the page for our project.  On the left is a Build Now option – let’s try it out…  After a few seconds, you should see a progress bar in the bottom left, telling you that Hudson is building your project.  If nothing happens, check the top right of the page for an Enable auto-refresh option.

Once finished, we’ll see a blue ball that indicates a successful build (your code in source control does build, right?).  If instead you see a red icon, then the build has broken, and we need to fix something.  Either way, click on the most recent build link to see details of the build:

On this page, we can see the status of this particular build (remember blue ball means success), the changes in source control from the last build (none in this case, as we manually triggered the build) and the console output, where we can see the details of the code being checked out from source control and built using XCode.  If your build has failed, this is a good place to look for the problem.

Phew!  The system works!  Checking Twitter, sure enough there’s a success message there:

But hang on, where’s my email?  The problem here is that we need to configure Hudson with an smtp server so that it can deliver mail to us.  Go back to the Manage Hudson page, and click on Configure System.  Near the bottom you can configure the global settings for email and Twitter.  If you’re unsure what smtp server to use, check the settings for your email program – you’re looking for the name of your outgoing smtp server.

So enter in your smtp details, and the system admin email address (that’s you).  Click on the Test button and make sure that you get the test email message.

Going back to the main Hudson page, you should notice that our project is listed, and on the right hand side is a Build Now icon (so you don’t have to go into the project to kick off a build).  Click it to start another build, and then click on your project to follow the build’s progress…  Still no email.  Bugger.  Going to the project’s Configure screen, I look up the help for email, which sheds some light on the issue:

This is sensible – no point in spamming out emails when builds are all going well.

So by now we know that we can check out our code and build it automatically.  But where is the build?  Hudson creates a working folder in your root directory.  Have a look in ~/.hudson, and you’ll see all of Hudson’s internals.  In particular, the folder ~/.hudson/jobs/Awesome Game contains all of our build history and a workspace folder in which the code is checked out and built.  So if I want to get the built app, I can grab it from:

~/.hudson/jobs/Amazing%20Game/workspace/build/Release-iphoneos/

Later we’ll look at packaging up the build and putting it somewhere handy, but for now, let’s make sure that it’ll pick up on changes in source control.  Check out a working copy (if you don’t have one) of your project and put in a small change that will cause the build to break.  This ought to do it:

Commit the change (or commit and push if you’re using a remote Mercurial server like I am) and wait for a minute.  You may notice a message mentioning a ‘quiet period’ before the build kicks off properly.  This quiet period is Hudson waiting after it detects a change in source control to make sure all changes have been committed – this is particularly useful with CVS, where every file is committed separately.

Ah, BOOM!  Our build has failed nicely.  And I got an email this time, which is nice  A broken build is  something that shouldn’t be left broken, so let’s fix it and check the build comes back happily.  Fix the code, push, commit, wait… success!  Twitter and email confirm everything’s OK.

We’ve covered quite a bit so far so it’s probably a good time to take a break and see what we’ve achieved, and how that matches our goals from last week.

What works?

• Monitor source control (Mercurial and Subversion in my case), kicking off a build whenever changes are checked in.
• We’ll be able to manually kick off a build as needed.
• The build will grab all our source code from source control and build it in all relevant configurations – debug, release, simulator, device and so on.
• When the build is finished, it’ll notify people of the success / failure.
• There’ll be a handy location (web page most likely) where we can monitor / control the whole thing.

What do we still need to do?

• We’ll schedule a nightly build as well, just in case anything outside of source control has changed. This is fairly simple – an exercise for the reader!
• Copy a successfully built app to somewhere useful, preferably packaging it up with a provisioning profile in a zip archive, ready for emailing.
• Automatically update version numbers on our ‘stable’ build with each build.
• Bundle up the resulting builds with a provisioning profile and copy them to a shared folder.
• Start up automatically under when your build machine is booted.
• Paint a unicorn.

OK, plenty to work on, I’ll be back next week with some more detail, or if you’re keen have a crack yourself and let me know how you get on!

Disclaimer: I’ll let you in on a secret now, to avoid disappointment later on.  By the end of this article you won’t have a fully working automated build  The reason is that I was half way through writing up this article using CruiseControl when I started to look at Hudson, which seems like it might be promising.  So this week is an introduction to the idea of continuous integration and automated builds and how they can save you time and grief.  Next week I’ll aim to have a useful build system up and running along with instructions on how to set it up…

What?  Why?

An automated build is a system that can build your software from source to finished product without human intervention.  At it’s simplest, this can be an XCode project or a makefile – after all, you don’t compile each file and link them together by hand, do you?  The key idea is that we can remove human error from the process, making it more efficient.  The ideal is to have a system whereby you can build all versions of your software with a single click (and also score a couple of points on the Joel test).

Consider the following situation.  You want a fresh build of the code to show off to your boss, “Reckless” Jim, to keep him off your back.  When you wake up, you grab your iPhone 4 (lucky bastard), open Safari and click a link on a web page before you get in the shower (remember to put the iPhone down first).  While you’re halfway through your hot water supply (and the Sound of Music soundtrack), the following is happening:

Cool, huh?

Or how about this:

Neat, eh?

This second example is what’s known as continuous integration – every time you check in new code to source control, the whole build is exercised to make sure nothing’s broken.  Why?  Often a broken build is a simple 30 second fix at the time, but 2 weeks later it might require hours of sorting through commit logs to see where the problem is.

The solution

So, what does it take to set up a system like this?  Not as much as you might think.  There are some neat tools out there to get you up and running fairly quickly.  The problem is that I didn’t have time to try out both of the likely candidates for this week’s deadline.  I’ve used CruiseControl before, and Hudson looks promising.

So here’s the goal for next week.  I’ll show you how to set up a system that will:

In the mean time here’s a couple of sample chapters from Pragmatic Project Automation: How to Build, Deploy, and Monitor Java Applications that covers the basics of setting up CruiseControl.

So, now to decide between Hudson and CruiseControl.  If you have any thoughts on either of these systems (or another I haven’t heard of), please let me know in the comments below…

Obtaining a Fix

The Pragmatic guys frequently run special promotions, and it’s quite common to get up to 50% off many titles.  In particular if you sign up and then recommend a friend to sign up, you’ll both immediately get 40% off coupons to redeem on any purchase.  If you don’t have any friends who are sufficiently book-nerdy like you and me, drop me a line at george.sealy ‘at’ gmail.com and I’ll send you an invite so we can both enjoy a discount

Some Great Books

The Art of Game Design : A Book of Lenses

Beginning iPhone 3 Development: Exploring the iPhone SDK

User Interface Design for Programmers

Game Engine Architecture

The Pragmatic Programmer: From Journeyman to Master

My Name is George

Some days it’s hard to think of something to write about, and other days something just falls in your lap. Happily this week it was the latter. It all started with Miguel (as these things often do) – a.k.a. @mysterycoconut on Twitter – posting this article about connecting to an iDevice via telnet while running your App. Using this technique it becomes possible to rapidly tweak parameters on the fly or in fact do all sorts of crazy stuff.

Needless to say I jumped on this idea to aid in tuning my current game prototype. I’ve just reached the stage where rapid tweaking of variables is going to help focus and refine the game play. Miguel suggests hooking a Lua scripting engine into your code, but I thought it would be fun to create my own, plus most of my code is C++ (rather than Objective-C) and I’ve found binding Lua to C++ to be a fiddly process in the past.

The corresponding code for this article can be found here. Note that the DebugServer code comes from Miguel’s article, and the AsyncSocket code comes from here.   I’ve added a minimal amount to Miguel’s code to pass received commands on to the DebugUtils functions found in Debug.[h|mm]. I haven’t had time to create a complete sample project, but the code provided should drop into any existing code base with minimal fuss. If the coding style looks a little strange, you may want to also read this and this.

So, let’s start at the end and work backwards, which I often find to be an effective technique. When tuning a game we spend a lot of time tweaking variables and constants – maximum speed, rate of fire, gravity, item prices, animation durations etc. Wouldn’t it be nice to open up a telnet session to your App and enter a command like “gravity = -8”? Alternatively, maybe there’s a set pattern of enemies that causes a rare bug. Wouldn’t it be nice to enter a “SetupWeirdEnemyPattern” command for easy testing?

So, we have two common scenarios – tweaking a variable and executing a function that alters the game state in a slightly more involved way. The code below allows us to do this. Variables and functions can be “registered” and given a name, which we can then use to adjust values or execute functions via our telnet connection. To make it simple to pass in parameters to these functions, we pass in a set of text parameters using argc and argv, just like the main() function of a C++ program.

Let’s have a look at the public interface to this code:

typedef bool (*DebugCommandFunc)(int argc, char *argv[]);

namespace DebugUtils
{
    void Initialise();

    bool ProcessCommand(const char *command);

    bool RegisterCommand(const char *commandName, DebugCommandFunc commandFunc);

    bool RegisterVariable(const char *variableName, bool *var);
    bool RegisterVariable(const char *variableName, int *var);
    bool RegisterVariable(const char *variableName, float *var);
}

Not much there, is there? The guts of it is in the ProcessCommand() function, where we pass in a string, and magic happens. Or, to provide an example:

bool ResetGameCommand(int argc, char *argv[])
{
// Do stuff here to reset the game state
}

DebugUtils::RegisterCommand(“ResetGame”, ResetGameCommand);

And later, through a telnet connection, we enter “ResetGame”, which in turn calls:

DebugUtils::ProcessCommand(“ResetGame”);

And, just like magic, we reset the game state.

Or as another example, we could register a game variable, say…

DebugUtils::RegisterVariable(“gold”, &gameState.pileOfDubloons);

And then, through the telnet connection, we can tweak this value mid-game, by entering:

“gold = 10000”, or “gold *= 2.0”

I’ve intentionally kept this system as simple as possible, but there’s a lot of power here. It’s very simple to register variables or add short pieces of code that can be run as needed. Just a few examples – turning on debug drawing to see bounding boxes of physics objects, changing the value of gravity or a jump pack’s thrust, loading an arbitrary level, turning off collision detection – the sky’s the limit.

The key thing here is to let you spend as much time as possible refining and tuning your game, because as Jesse Schell describes in his excellent book The Art of Game Design: A book of lenses:

The Rule of the Loop: The more times you test and improve your design, the better your game will be.

In other words, by removing the code-compile-run stage from tuning our game, we can create a better game in less time. Hopefully this code, combined with Miguel’s will provide a simple way to help you refine and debug your game.  Happy tuning!

P.S. I was planning to include a description of how I used TDD (Test Driven Development), but figured this post was long enough. If it’s something you’re interested in though, leave a comment or drop me a note on Twitter (@GeorgeSealy) and I’ll put together an article.

P.P.S. Yes, there’s a few things that could be added to this code to make it even more useful. At the moment it’s not possible to pass a parameter that’s a string with spaces or new lines in it. To do this, you just extend the tokenizing code to respect quoted parameters like “This is a string”.

P.P.S. I’ve resisted the temptation to allow more complex calculations such as “percentComplete = 100 * itemsFound / totalItems”. Most of the time all you want to do is tweak values. If you need this much power, hook in Lua as Miguel suggests, don’t roll your own as it becomes increasingly difficult to maintain as a project matures.

P.P.P.S. I’ve also just realised one of the more useful things I’ve left out – if you just supply the name of a variable on its own, then the variable’s value should be either returned via telnet or printed out using NSLog or similar.  There’s nothing worse than finding the perfect value for something and then being unable to see what it is!

P.P.P.P.S. The observant amongst you may have noticed that the link to Jesse Schell’s book is an Amazon Affiliate link, meaning if you follow that link and subsequently purchase something I may get a small commission. I will only ever put such links to books that I have bought myself or at least read cover to cover and would recommend.

If you’re an iPhone Dev, and you’re not using Twitter, you really should be.

Whatever else Twitter may be, it is a wonderful, anarchic repository for random bits of incredibly useful information. A good habit to get into is marking useful tidbits as favourites so that you can come back to them later. (As an aside, the first person who comes up with an easy way to sync up favourites on Twitter, bookmarks from various browsers on different machines and favourites in Google’s RSS reader will do very well for themselves.)

In celebration of the randomness of twitter, here’s a few really handy iPhone-related tips I’ve picked up lately:

Managing double size graphics for the iPhone 4 (link)

With the release of the iPhone 4, developers now have the tricky problem of dealing with multiple resolutions. There seems to be a consensus amongst many iPhone devs that the best way to handle this is with 2 sets of graphics, one for each resolution. So this means you have to deal with two sets of filenames through out your code, right? Well actually no.

It turns out that when you call:

[UIImage imageNamed:@"myImage.png"]; (or contentsOfFile:)

on the iPhone 4, it will actually first attempt to load an image called myImage@2x.png first.  So simply ensure that both sets of images have the same file names (with @2x added to the iPhone 4 set), and the Sdk will take care of things for you.

Update: It turns out that the ‘@’ symbol can cause havoc with Subversion (SVN).  If you use SVN, you’d better read this before you start creating all those double resolution graphics.  Once again, I learned this from Twitter also.

Clear the colour buffer for faster rendering (link via @majicDave)

Because of the way iOS handles OpenGL rendering, it’s possible to get a speed boost by ensuring you make a call to clear the colour buffer before rendering. This is actually faster than not making the call at all, which seems counter-intuitive.  @majicDave reported a significant improvement in frame rate when he tried this.

Hot loading levels on the device (link)

If you have a chance, read @SnappyTouch’s article in Game Programming Gems 6 : Stay in the Game: Asset Hotloading for Fast Iteration. The idea is to allow assets (textures, levels, shaders, AI scripts etc) to be re-loaded in game.

By doing this, it becomes possible to rapidly tweak and refine your game without the need to recompile. One obvious benefit is that refining assets becomes far quicker without the necessary compile – and the faster you can iterate, the better your result will be. A second benefit is that it becomes possible for artists and designers to tinker with the look and game play of the App without needing to involve a programmer.

Hot loading assets on an iPhone / iPod is a little more involved, as we don’t get easy access to the device’s file system via another computer.  Miguel’s article demonstrates a good way to do this. This was also the article that kicked off the idea of #iDevBlogADay.

Getting old firmware versions (link)

So you realize that you need to test an odd bug in OS 3.13 but you’ve just upgraded your testing device to OS 4.0? This tweet give you all you need to know to revert your device to an earlier OS version. Just remember to ensure you back up everything important!

There’s HOW many sizes of icons? (link)

‘Nuff said.

Tracking your day to day sales on the App Store? (link)

Admittedly Apple emailed all developers about this one, but I heard it first via Twitter.

How to handle special characters in URLs (link)

So you’re dealing with URLs and want to get rid of all those pesky escape characters like ‘%20’ (a space)? Don’t do this by hand, it’s messy, error prone and just not needed. Apple have given us stringByReplacingPercentEscapes to take care of this for us.  This is the kind of thing that can be found in documentation if you know to look for it.  Often a quick question asked on Twitter can gain valuable insight.

Physics calculations are tricky (link)

Fortunately there’s clever people out there who can help us avoid the most common errors. By fixing the time step at which you run your simulation, many problems can be avoided. Additionally, there can be many advantages to separating the render and simulation time steps. http://gafferongames.com provides a wealth of useful information here.

I’ve been surprised by Twitter, I was worried it was just another social networking Web 2.0 buzz annoyance, but I was lucky to find a vibrant, fun group of intelligent people with a common interest.  If you’re not sure about Twitter, give it a go.  I’m @GeorgeSealy, feel free to say hi – I promise to keep my tweets to mainly iPhone Dev related topics!

This site has been quiet for too long.  That’s why I’ve got involved with the #iDevBlogADay initiative.  There’s a bunch of great indie developers contributing as part of a regular roster.  If you want to follow all of us just subscribe to this feed.

Any good programmer is always looking for new tools and ideas to improve the code they write and to become more productive. For myself, fifteen years after I began coding professionally, this is truer than ever.

So when Noel Llopis started dropping odd comments on Twitter about ‘no classes’, ‘plain old data structures’ and ‘no singletons’ I was intrigued. How did this work? Where did this style come from? Did it lead to simple, effective code? Was it a style that would work for me? Is it a style that will work for you?

If you haven’t already read Noel’s article from yesterday describing his current coding style, then go ahead and read it now.

Take your time, I’ll wait.

It’s definitely an unusual style, and flies in the face of ‘traditional’ wisdom. So what’s life like in a world without objects?  Pretty cool, actually. I’ve been experimenting with it for the last couple of weeks, and here are a few observations that might encourage you to give this style a go.

The ‘Rules’

It’s hard to change style so dramatically, so a few rules of thumb are useful to help keep on track.  This style is sufficiently different that it takes time to get your head around it.  Prepare to spend some time writing (and re-writing) your first code this way.

• No classes.
• Use simple structs to store data / state information, making everything visible.
• Avoid dynamic memory allocation and pointers in general.
• Nothing is global – all data is passed as needed to functions that require it.
• Functions have a single purpose – for a given set of input data, generate an output, avoiding any side effects.
• Use TDD for everything.

TDD Becomes Much Easier

Test Driven Development (TDD) has always appealed to me, but to be honest my history with TDD is a bit like my attempts to build a model railway – plenty of well intentioned starts but no completed projects. Sooner or later I’d always give up on TDD because it just didn’t fit the way I’m used to coding.

But in a world of simple data and functions with a single purpose, TDD comes into it’s own. The tests you write drive the code to be as simple as possible with none of the cruft that typically accumulates in class definitions. And constructing tests is simple – create some simple test data, pass it into a function and check the resulting outputs. All tests follow this simple formula, with not a single pesky mock object in sight.

Lean Code And A Sense Of Progress

The resulting code feels resistant to bloat. Nothing is hidden, and every function has a clear purpose with no unexpected side effects. Time will tell if this pattern persists as the code matures.  I’m constantly surprised by the way the code evolves down unexpected avenues.

Every test, every function has a purpose that is clearly defined.  This simplicity has a huge appeal. A test and the code that goes with it can be written fairly quickly, and each additional test provides visible progress.

Struct Is The New Interface

I’m used to relying heavily on interfaces or abstract classes in my coding, and the lack of such an option concerned me somewhat. It still does – I miss the option to have a collection of ‘game objects’ and know that I can call Update() on all of them no matter what they are – Crates, Zombies or Flying Quesadillas.

What did surprise me was that rather than accumulating everything into a single class, the data naturally falls into a collection of several parts more in keeping with the idea of components in an OO system (composition over inheritance). For example, when I was writing a leaf particle system the resulting code combined the following bits of data – a mesh, an array of (generic) particles, an array of leaf specific particle data and a free-list for deleting and creating particles from a statically allocated array.

This fragmentation may seem unfortunate, but in fact it’s a strength. The functions that deal with a mesh are applicable to all meshes, not just rendering particles. The free-list struct and functions can be used anywhere that elements in an array need to be frequently created and destroyed.  Where in an OO design you carry around an entire object to each method or function call, with this approach you only pass around the ‘part’ of the object that is directly relevant.

Coding Is Fun Again

Ultimately this is the most important point for me. I’m enjoying coding this way. Progress is rapid, and through the use of unit testing, progress is very visible.  My free time for game programming is extremely limited, so the ability to write code in small, manageable chunks is a big win.  Refactoring can be undertaken with confidence, knowing that the unit tests and the basic structure of the code give some protection against unintended side effects.

This style won’t suit everyone, and it will absolutely terrify the Architecture Astronauts, but if you’re interested then give it a go and let me know how you get on!

Regardless of whether you consider Apple to be Good or Evil (or perhaps just a business who wants to make money?), the latest changes to the iPhone Developers agreement make most developers wonder what their back up plan is.

One of our goals at Acorn Heroes has always been to keep our code as portable as possible.  By doing this, if/when we find ourselves needing to move code over to another platform (for whatever reason), the process is not too cumbersome.  As such, C++ makes a natural, portable choice, right?

Well, sorta.  We write a collection of C++ code that we can carry over to, say, the PC.  The problem here is that so much useful code for the iPhone ( loading images, accelerometer, touch screen, geo-location etc) is based in Objective-C.  Objective-C code isn’t supported on many platforms at all.  So despite being able to write hybrid C++/Obj-C code  we’re pretty much screwed, right?

Not really.  The key realization to make is that, on any platform, even if we use the same language, we’ll end up re-writing most of our code that interacts with the OS anyway – things like getting the system time, creating windows and loading files will always be specific to our chosen platform.

What we don’t want though, is for these platform changes to affect the rest of our code base.  Here’s how it can be done:

Make all your header files pure C++.

It’s that simple.  As an example, consider a class that wraps an OpenGL texture.  We create an instance of this class, supplying the filename of the texture as a parameter to the constructor.  The class may have enable / disable methods, and perhaps a few other bits and pieces like making the texture wrap, or perhaps looking up pixel data.  The header here is pure C++.

The implementation however, is most likely to be heavily littered with Objective-C for things such as getting a system path to our data folder, loading the file and binding a PVRTC texture.  And that’s OK.

So why is this such a great way to do things?

• We’ve implemented our code in a way that’s best suited for a given platform.  It’s simpler to code and is probably more efficient or optimal than a generic, cross-platform solution.
• Our code isn’t littered with #ifdefs supplying different code paths for different operating systems.
• The bulk of our code, especially the all important game logic, is completely portable – at least as portable as any other language out there (yes Java, I’m looking at you).
• We haven’t wasted large amounts of effort writing a code base that can be cross-compiled for multiple platforms.  Doing this needs to be tackled early in a project, meaning it sucks away our time at a stage when we should be concentrating on defining our game play.
• When we sell 1,000,000+ copies and have the urge to go cross-platform, we take the existing code base pull out a well defined set of implementation files and recode them for the new platform.  Easy, and we would need to re-write that code anyway.

We’re currently prototyping some ideas for our next game. This time round we’re looking to do something fun with light and dark. Unfortunately the fixed OpenGL pipeline has a fairly standard look to it, and we can’t move to using shaders in OpenGL 2 without alienating a large number of people (including ourselves – our personal devices are both iPod Touches).

So as well as game play ideas, we’ve been prototyping the tech we’ll need to bring our world to life. The game will be 2d, but we really want to find a good way to render lights and shadows to create a dark, threatening experience.

After some Google-ing I found a good starting point over on GameDev.net. The basic idea is that each light is rendered into a buffer (the alpha buffer in this case), and then any shadows are rendered on top, blacking out the light. The alpha buffer is then combined with the scene geometry to add colour to the scene.

Fairly quickly I was able to get the basic algorithm up and running. The problem though is that each light requires rendering a screen’s worth of data, and frame rates drop drastically after adding more than 2 or 3 lights. So I started looking around for other options.

One idea that appealed was to render all the lights into a single buffer, and then use a single pass to lay this back over the scene. The problem here is that we can’t just render each light’s shadows into the buffer as black.

The answer here is to calculate the shadows for a light and incorporate them into the mesh used to render the light – i.e. instead of, say, a circular triangle fan for a light source, we generate a mesh for the circular fan minus the shadows.

So, after a few false starts I managed to get the basic algorithm working, and you can see its development in the screen shots on this page. The good news is that we can have large (double digit) numbers of lights in a scene, all casting shadows and all able to be moved or turned on and off dynamically each frame. Currently performance is about 45fps on a 2nd gen iPod Touch, but I’m fairly confident that can be drastically improved with some basic optimisation.

Once I’ve got the code properly under control, I’ll write an article covering the technique in some depth, if people are interested.

The last screen shot here shows the algorithm working at the moment. I still need to add spot lights rather than point light sources, but that should be a fairly simple thing to add. There’s a couple of bugs still to iron out too, but I think it proves the concept well enough to let us move onto prototyping other parts of the game. The addition of textures makes a huge difference too. While this is still very rough, I think there’s some real promise here, and can’t wait to start building the game that makes the most of it!

Beg.  Borrow.  Steal.

Got an existing game, and you feel comfortable with the code?  Take a copy, rip out the guts of it, and start prototyping.  Need to perform a Fourier transform?  Do a Google search and mash the code you find into your prototype.  Know that a friend has solved a given problem before?  Ask them if they can send you the relevant code.

So you’ve found some useful code, but it doesn’t represent vectors the same way you do?  Don’t go through the borrowed code doing a search/replace – it’ll take too long and may introduce errors.  Instead, simply convert vector data on the way in and out with a simple piece of wrapper code.  Remember, CPU and memory are free.

Rapid prototyping is about getting results fast.  Sometimes the fastest way to get a result is to use working code from somewhere else.